Context Modeling for LLM-Based Behavioral Simulation

ACL ARR 2026 January Submission3107 Authors

Keywords: LLM-based simulation, Conditional generation, Decision-making

Abstract: Large language models (LLMs) offer new opportunities for simulating human decision making at scale, yet prompting-based approaches often fail to recover realistic population-level response distributions, particularly for heterogeneous demographic groups and high-entropy questions. This paper introduces context modeling with agentic AI, a structured simulation framework that explicitly constructs latent, individual-level decision contexts prior to response generation. By separating persona construction from decision making and introducing structured stochasticity through context sampling, the method captures within-group heterogeneity and mitigates mode collapse inherent in single-prompt simulations. Extensive empirical evaluation across population-level and group-conditioned settings shows that context modeling consistently improves simulation fidelity, with especially strong gains on high-entropy questions and demographic subpopulations, while preserving performance on low-entropy tasks. Detailed tail analysis further reveals that the primary improvement arises from recovering minority and low-probability response options that are systematically underrepresented by direct prompting. Overall, the results demonstrate that faithful behavioral simulation requires explicit modeling of contextual heterogeneity and structured uncertainty, establishing context modeling as a robust foundation for LLM-based behavioral simulation.

Paper Type: Long

Research Area: Computational Social Science, Cultural Analytics, and NLP for Social Good

Research Area Keywords: Language Modeling, Interpretability and Analysis of Models for NLP

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 3107